In a change of course, Microsoft is to start working more closely with Facebook on open-source AI. The two companies had previously been competing, with Cognitive Toolkit and PyTorch respectively, and both failing to keep up with the adoption of Google TensorFlow.
Since its launch in 2015, Google TensorFlow has been the most popular open-source AI software platform for developers.
The library was put together by the Google Brain team and provides tools for numerical computation and machine learning at scale. Countless machine and deep learning models and algorithms are brought together in an accessible format, with help from Python and C++.
The challenge for competitors is to match both the depth of the resource and the ease of access.
The open-source approach
One way of doing that is to work together, which explains Microsoft’s recent shift toward supporting Facebook in the open-source AI space, rather than competing with the social media giant.
It’s certainly not a complete admission of defeat from Microsoft, though. The company will continue to support Cognitive Toolkit. But user behaviour doesn’t lie, as Eric Boyd, Microsoft’s corporate vice president of AI platform explains:
The momentum of community, really, is around PyTorch and TensorFlow, and so that’s where we’re throwing the bulk of our emphasis.
“Having community adoption is hugely important.”
Internet of Business says
Microsoft has never shied away from backing software and hardware from rivals, instead of focusing solely on the company’s own developments. With Google well out in front when it comes to open-source AI, Microsoft’s decision to join forces with, and quietly contribute to, Facebook’s PyTorch platform makes a lot of sense.
Microsoft’s answer to TensorFlow, Cognitive Toolkit (CNTK) was released in 2016. The statistics below – from Google researcher and creator of AI toolkit Keras, Francois Chollet – suggest that, in the field of AI research at least, Microsoft’s platform is lacking a substantial base of academic users.
TensorFlow is the platform of choice for deep learning in the research community. These are deep learning framework mentions on arXiv over the past 3 months pic.twitter.com/v6ZEi63hzP
— François Chollet (@fchollet) March 8, 2018
Therefore, it’s no surprise that Microsoft has stepped back from its own offering. Although Cognitive Toolkit is well regarded for its ability to build speech recognition systems, Facebook’s PyTorch has seen wider adoption.
In 2017 Microsoft and Facebook concluded that the AI ecosystem was overcrowded. Their solution, in an attempt to “defrag some of the complexity”, was ONNX, a software platform that simplifies the process of exporting models trained with one AI framework to another.
With so many AI frameworks out there, and Google dominating the space, Microsoft and Facebook are wise to consolidate their efforts, though they would have preferred to have found greater success independently.
It’s also important to acknowledge the commitment to open-source software in the field, which Microsoft has been at pains to reaffirm its affection for.